setup.py 12 KB
Newer Older
Tri Dao's avatar
Tri Dao committed
1
2
3
4
# Adapted from https://github.com/NVIDIA/apex/blob/master/setup.py
import sys
import warnings
import os
5
6
import re
import ast
Tri Dao's avatar
Tri Dao committed
7
from pathlib import Path
Tri Dao's avatar
Tri Dao committed
8
from packaging.version import parse, Version
9
import platform
Tri Dao's avatar
Tri Dao committed
10
11
12
13

from setuptools import setup, find_packages
import subprocess

Pierce Freeman's avatar
Pierce Freeman committed
14
15
import urllib.request
import urllib.error
Tri Dao's avatar
Tri Dao committed
16
17
import torch
from torch.utils.cpp_extension import BuildExtension, CppExtension, CUDAExtension, CUDA_HOME
18
from wheel.bdist_wheel import bdist_wheel as _bdist_wheel
Tri Dao's avatar
Tri Dao committed
19
20
21
22
23


with open("README.md", "r", encoding="utf-8") as fh:
    long_description = fh.read()

Tri Dao's avatar
Tri Dao committed
24
25
26
27

# ninja build does not work unless include_dirs are abs path
this_dir = os.path.dirname(os.path.abspath(__file__))

28
PACKAGE_NAME = "flash_attn_wheels"
Tri Dao's avatar
Tri Dao committed
29

30
31
32
33
34
35
36
37
38
# @pierce - TODO: Update for proper release
BASE_WHEEL_URL = "https://github.com/piercefreeman/flash-attention/releases/download/{tag_name}/{wheel_name}"

# FORCE_BUILD: Force a fresh build locally, instead of attempting to find prebuilt wheels
# SKIP_CUDA_BUILD: Intended to allow CI to use a simple `python setup.py sdist` run to copy over raw files, without any cuda compilation
FORCE_BUILD = os.getenv("FLASH_ATTENTION_FORCE_BUILD", "FALSE") == "TRUE"
SKIP_CUDA_BUILD = os.getenv("FLASH_ATTENTION_SKIP_CUDA_BUILD", "FALSE") == "TRUE"


39
40
def get_platform():
    """
41
    Returns the platform name as used in wheel filenames.
42
43
44
45
    """
    if sys.platform.startswith('linux'):
        return 'linux_x86_64'
    elif sys.platform == 'darwin':
46
47
        mac_version = '.'.join(platform.mac_ver()[0].split('.')[:2])
        return f'macosx_{mac_version}_x86_64'
48
49
50
51
52
53
    elif sys.platform == 'win32':
        return 'win_amd64'
    else:
        raise ValueError('Unsupported platform: {}'.format(sys.platform))


Tri Dao's avatar
Tri Dao committed
54
55
56
57
def get_cuda_bare_metal_version(cuda_dir):
    raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True)
    output = raw_output.split()
    release_idx = output.index("release") + 1
Tri Dao's avatar
Tri Dao committed
58
    bare_metal_version = parse(output[release_idx].split(",")[0])
Tri Dao's avatar
Tri Dao committed
59

Tri Dao's avatar
Tri Dao committed
60
    return raw_output, bare_metal_version
Tri Dao's avatar
Tri Dao committed
61
62
63


def check_cuda_torch_binary_vs_bare_metal(cuda_dir):
Tri Dao's avatar
Tri Dao committed
64
65
    raw_output, bare_metal_version = get_cuda_bare_metal_version(cuda_dir)
    torch_binary_version = parse(torch.version.cuda)
Tri Dao's avatar
Tri Dao committed
66
67
68
69

    print("\nCompiling cuda extensions with")
    print(raw_output + "from " + cuda_dir + "/bin\n")

Tri Dao's avatar
Tri Dao committed
70
    if (bare_metal_version != torch_binary_version):
Tri Dao's avatar
Tri Dao committed
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
        raise RuntimeError(
            "Cuda extensions are being compiled with a version of Cuda that does "
            "not match the version used to compile Pytorch binaries.  "
            "Pytorch binaries were compiled with Cuda {}.\n".format(torch.version.cuda)
            + "In some cases, a minor-version mismatch will not cause later errors:  "
            "https://github.com/NVIDIA/apex/pull/323#discussion_r287021798.  "
            "You can try commenting out this check (at your own risk)."
        )


def raise_if_cuda_home_none(global_option: str) -> None:
    if CUDA_HOME is not None:
        return
    raise RuntimeError(
        f"{global_option} was requested, but nvcc was not found.  Are you sure your environment has nvcc available?  "
        "If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, "
        "only images whose names contain 'devel' will provide nvcc."
    )


def append_nvcc_threads(nvcc_extra_args):
Tri Dao's avatar
Tri Dao committed
92
93
    _, bare_metal_version = get_cuda_bare_metal_version(CUDA_HOME)
    if bare_metal_version >= Version("11.2"):
Tri Dao's avatar
Tri Dao committed
94
95
96
97
98
99
100
101
102
103
104
        return nvcc_extra_args + ["--threads", "4"]
    return nvcc_extra_args


if not torch.cuda.is_available():
    # https://github.com/NVIDIA/apex/issues/486
    # Extension builds after https://github.com/pytorch/pytorch/pull/23408 attempt to query torch.cuda.get_device_capability(),
    # which will fail if you are compiling in an environment without visible GPUs (e.g. during an nvidia-docker build command).
    print(
        "\nWarning: Torch did not find available GPUs on this system.\n",
        "If your intention is to cross-compile, this is not an error.\n"
Tri Dao's avatar
Tri Dao committed
105
106
        "By default, Apex will cross-compile for Pascal (compute capabilities 6.0, 6.1, 6.2),\n"
        "Volta (compute capability 7.0), Turing (compute capability 7.5),\n"
Tri Dao's avatar
Tri Dao committed
107
108
109
110
        "and, if the CUDA version is >= 11.0, Ampere (compute capability 8.0).\n"
        "If you wish to cross-compile for a single specific architecture,\n"
        'export TORCH_CUDA_ARCH_LIST="compute capability" before running setup.py.\n',
    )
Tri Dao's avatar
Tri Dao committed
111
112
113
114
115
116
117
118
    if os.environ.get("TORCH_CUDA_ARCH_LIST", None) is None and CUDA_HOME is not None:
        _, bare_metal_version = get_cuda_bare_metal_version(CUDA_HOME)
        if bare_metal_version >= Version("11.8"):
            os.environ["TORCH_CUDA_ARCH_LIST"] = "6.0;6.1;6.2;7.0;7.5;8.0;8.6;9.0"
        elif bare_metal_version >= Version("11.1"):
            os.environ["TORCH_CUDA_ARCH_LIST"] = "6.0;6.1;6.2;7.0;7.5;8.0;8.6"
        elif bare_metal_version == Version("11.0"):
            os.environ["TORCH_CUDA_ARCH_LIST"] = "6.0;6.1;6.2;7.0;7.5;8.0"
Tri Dao's avatar
Tri Dao committed
119
        else:
Tri Dao's avatar
Tri Dao committed
120
121
            os.environ["TORCH_CUDA_ARCH_LIST"] = "6.0;6.1;6.2;7.0;7.5"

Tri Dao's avatar
Tri Dao committed
122
123
124
cmdclass = {}
ext_modules = []

125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
if not SKIP_CUDA_BUILD:
    print("\n\ntorch.__version__  = {}\n\n".format(torch.__version__))
    TORCH_MAJOR = int(torch.__version__.split(".")[0])
    TORCH_MINOR = int(torch.__version__.split(".")[1])

    # Check, if ATen/CUDAGeneratorImpl.h is found, otherwise use ATen/cuda/CUDAGeneratorImpl.h
    # See https://github.com/pytorch/pytorch/pull/70650
    generator_flag = []
    torch_dir = torch.__path__[0]
    if os.path.exists(os.path.join(torch_dir, "include", "ATen", "CUDAGeneratorImpl.h")):
        generator_flag = ["-DOLD_GENERATOR_PATH"]

    raise_if_cuda_home_none("flash_attn")
    # Check, if CUDA11 is installed for compute capability 8.0
    cc_flag = []
    _, bare_metal_version = get_cuda_bare_metal_version(CUDA_HOME)
    if bare_metal_version < Version("11.0"):
        raise RuntimeError("FlashAttention is only supported on CUDA 11 and above")
    cc_flag.append("-gencode")
    cc_flag.append("arch=compute_75,code=sm_75")
Tri Dao's avatar
Tri Dao committed
145
    cc_flag.append("-gencode")
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
    cc_flag.append("arch=compute_80,code=sm_80")
    if bare_metal_version >= Version("11.8"):
        cc_flag.append("-gencode")
        cc_flag.append("arch=compute_90,code=sm_90")

    subprocess.run(["git", "submodule", "update", "--init", "csrc/flash_attn/cutlass"])
    ext_modules.append(
        CUDAExtension(
            name="flash_attn_cuda",
            sources=[
                "csrc/flash_attn/fmha_api.cpp",
                "csrc/flash_attn/src/fmha_fwd_hdim32.cu",
                "csrc/flash_attn/src/fmha_fwd_hdim64.cu",
                "csrc/flash_attn/src/fmha_fwd_hdim128.cu",
                "csrc/flash_attn/src/fmha_bwd_hdim32.cu",
                "csrc/flash_attn/src/fmha_bwd_hdim64.cu",
                "csrc/flash_attn/src/fmha_bwd_hdim128.cu",
                "csrc/flash_attn/src/fmha_block_fprop_fp16_kernel.sm80.cu",
                "csrc/flash_attn/src/fmha_block_dgrad_fp16_kernel_loop.sm80.cu",
            ],
            extra_compile_args={
                "cxx": ["-O3", "-std=c++17"] + generator_flag,
                "nvcc": append_nvcc_threads(
                    [
                        "-O3",
                        "-std=c++17",
                        "-U__CUDA_NO_HALF_OPERATORS__",
                        "-U__CUDA_NO_HALF_CONVERSIONS__",
                        "-U__CUDA_NO_HALF2_OPERATORS__",
                        "-U__CUDA_NO_BFLOAT16_CONVERSIONS__",
                        "--expt-relaxed-constexpr",
                        "--expt-extended-lambda",
                        "--use_fast_math",
                        "--ptxas-options=-v",
                        "-lineinfo"
                    ]
                    + generator_flag
                    + cc_flag
                ),
            },
            include_dirs=[
                Path(this_dir) / 'csrc' / 'flash_attn',
                Path(this_dir) / 'csrc' / 'flash_attn' / 'src',
                Path(this_dir) / 'csrc' / 'flash_attn' / 'cutlass' / 'include',
            ],
        )
Tri Dao's avatar
Tri Dao committed
192
193
    )

194
195
196
197
198
199
200
201
202
203
def get_package_version():
    with open(Path(this_dir) / "flash_attn" / "__init__.py", "r") as f:
        version_match = re.search(r"^__version__\s*=\s*(.*)$", f.read(), re.MULTILINE)
    public_version = ast.literal_eval(version_match.group(1))
    local_version = os.environ.get("FLASH_ATTN_LOCAL_VERSION")
    if local_version:
        return f"{public_version}+{local_version}"
    else:
        return str(public_version)

204

205
206
207
208
209
210
class CachedWheelsCommand(_bdist_wheel):
     """
     The CachedWheelsCommand plugs into the default bdist wheel, which is ran by pip when it cannot
     find an existing wheel (which is currently the case for all flash attention installs). We use
     the environment parameters to detect whether there is already a pre-built version of a compatible
     wheel available and short-circuits the standard full build pipeline.
211

212
213
     """
     def run(self):
214
        if FORCE_BUILD:
Pierce Freeman's avatar
Pierce Freeman committed
215
            return super().run()
216
217
218
219
220
221
222
223
224
225
226
227
228

        raise_if_cuda_home_none("flash_attn")

        # Determine the version numbers that will be used to determine the correct wheel
        _, cuda_version_raw = get_cuda_bare_metal_version(CUDA_HOME)
        torch_version_raw = parse(torch.__version__)
        python_version = f"cp{sys.version_info.major}{sys.version_info.minor}"
        platform_name = get_platform()
        flash_version = get_package_version()
        cuda_version = f"{cuda_version_raw.major}{cuda_version_raw.minor}"
        torch_version = f"{torch_version_raw.major}.{torch_version_raw.minor}.{torch_version_raw.micro}"

        # Determine wheel URL based on CUDA version, torch version, python version and OS
229
        wheel_filename = f'{PACKAGE_NAME}-{flash_version}+cu{cuda_version}torch{torch_version}-{python_version}-{python_version}-{platform_name}.whl'
230
231
232
233
234
235
236
237
        wheel_url = BASE_WHEEL_URL.format(
            tag_name=f"v{flash_version}",
            wheel_name=wheel_filename
        )
        print("Guessing wheel URL: ", wheel_url)
        
        try:
            urllib.request.urlretrieve(wheel_url, wheel_filename)
238
239
240
241
242
243
244
245
246
247
248
249
250

            # Make the archive
            # Lifted from the root wheel processing command
            # https://github.com/pypa/wheel/blob/cf71108ff9f6ffc36978069acb28824b44ae028e/src/wheel/bdist_wheel.py#LL381C9-L381C85
            if not os.path.exists(self.dist_dir):
                os.makedirs(self.dist_dir)

            impl_tag, abi_tag, plat_tag = self.get_tag()
            archive_basename = f"{self.wheel_dist_name}-{impl_tag}-{abi_tag}-{plat_tag}"
        
            wheel_path = os.path.join(self.dist_dir, archive_basename + ".whl")
            print("Raw wheel path", wheel_path)
            os.rename(wheel_filename, wheel_path)
251
252
253
        except urllib.error.HTTPError:
            print("Precompiled wheel not found. Building from source...")
            # If the wheel could not be downloaded, build from source
254
            super().run()
255
256


Tri Dao's avatar
Tri Dao committed
257
setup(
258
    # @pierce - TODO: Revert for official release
259
    name=PACKAGE_NAME,
260
    version=get_package_version(),
Tri Dao's avatar
Tri Dao committed
261
262
263
    packages=find_packages(
        exclude=("build", "csrc", "include", "tests", "dist", "docs", "benchmarks", "flash_attn.egg-info",)
    ),
264
265
266
267
268
    #author="Tri Dao",
    #author_email="trid@stanford.edu",
    # @pierce - TODO: Revert for official release
    author="Pierce Freeman",
    author_email="pierce@freeman.vc",
Tri Dao's avatar
Tri Dao committed
269
270
271
    description="Flash Attention: Fast and Memory-Efficient Exact Attention",
    long_description=long_description,
    long_description_content_type="text/markdown",
272
273
    #url="https://github.com/HazyResearch/flash-attention",
    url="https://github.com/piercefreeman/flash-attention",
Tri Dao's avatar
Tri Dao committed
274
275
    classifiers=[
        "Programming Language :: Python :: 3",
276
        "License :: OSI Approved :: BSD License",
Phil Wang's avatar
Phil Wang committed
277
        "Operating System :: Unix",
Tri Dao's avatar
Tri Dao committed
278
    ],
Tri Dao's avatar
Tri Dao committed
279
    ext_modules=ext_modules,
280
    cmdclass={
281
        'bdist_wheel': CachedWheelsCommand,
282
283
        "build_ext": BuildExtension
    } if ext_modules else {
284
        'bdist_wheel': CachedWheelsCommand,
285
    },
Gustaf's avatar
Gustaf committed
286
287
288
289
    python_requires=">=3.7",
    install_requires=[
        "torch",
        "einops",
Pavel Shvets's avatar
Pavel Shvets committed
290
        "packaging",
291
        "ninja",
Gustaf's avatar
Gustaf committed
292
    ],
Tri Dao's avatar
Tri Dao committed
293
)